Audit Data Integrity
Inspect existing code, contracts, manifests, and logs; do not recompute missing
results. Report each applicable check as pass, flaw, or unverified, with a
specific evidence path or exact missing evidence.
Checks
- Establish the semantic meaning and order of every source axis. Trace each
transpose, reshape, ravel, flatten, stack, concatenation, cache, and
reload that can change sample identity. Shape equality alone is insufficient.
- Verify that every feature row remains aligned with its label, subject,
condition, session, bin, and other grouping identifiers. Look for explicit
value-level assertions on representative samples.
- Verify train/test independence at the real sampling unit. Check subject,
session, site, family, temporal, repeated-measure, and augmentation leakage as
applicable.
- Verify that imputation, scaling, feature selection, dimensionality reduction,
resampling, threshold selection, and model selection are fitted only within
training folds.
- Establish the value domain and transform contract. Confirm that training,
exported artifacts, manifests, and inference use the same clipping,
missing-value handling, transform, and feature order.
For a bounded parallel review, give a method-reviewer the data contract,
pipeline files, and validation evidence. Ask for evidence and candidate findings,
not a verdict.
1---2name: audit-data-integrity3description: Audit scientific data semantics, sample and label alignment, leakage, group splits, preprocessing boundaries, and train-to-inference transforms. Use for any result based on datasets, feature matrices, tensors, repeated observations, or learned preprocessing.4---56# Audit Data Integrity78Inspect existing code, contracts, manifests, and logs; do not recompute missing9results. Report each applicable check as `pass`, `flaw`, or `unverified`, with a10specific evidence path or exact missing evidence.1112## Checks13141. Establish the semantic meaning and order of every source axis. Trace each15 `transpose`, `reshape`, `ravel`, `flatten`, `stack`, concatenation, cache, and16 reload that can change sample identity. Shape equality alone is insufficient.172. Verify that every feature row remains aligned with its label, subject,18 condition, session, bin, and other grouping identifiers. Look for explicit19 value-level assertions on representative samples.203. Verify train/test independence at the real sampling unit. Check subject,21 session, site, family, temporal, repeated-measure, and augmentation leakage as22 applicable.234. Verify that imputation, scaling, feature selection, dimensionality reduction,24 resampling, threshold selection, and model selection are fitted only within25 training folds.265. Establish the value domain and transform contract. Confirm that training,27 exported artifacts, manifests, and inference use the same clipping,28 missing-value handling, transform, and feature order.2930For a bounded parallel review, give a `method-reviewer` the data contract,31pipeline files, and validation evidence. Ask for evidence and candidate findings,32not a verdict.
neuroaihub